GraphSynth: Synthesis of Network Configuration Templates Using Large Language Models
X X Zhang, Xianming Gao, Peilin Tao, Tao Feng · 2025
We investigate whether Large Language Models can synthesize correct network configuration templates in a simpler and more efficient way with limited prompts. The experimental results demonstrate that the LLM's shortcomings in command relation understanding can be supplemented by external models to generate promising templates. We do not rule out the contribution of validators to property-specific consistency. However, our framework aims to preserve the large language model generation capability while focusing on the implicit relationships in configurations. By introducing a hierarchical command graph, we follow a top-down matching to generate templates, reducing feedback interactions in generation and bring potential solutions for automated provisioning. We process template generation under different scenarios, including establishing BGP basic connections on multiple routers and imposing restrictions such as route redistribution at specific locations. Human involvement is restricted to providing critical information. Experiment shows that expected templates can be generated in the first iteration, particularly for the basic task. The advanced task is influenced by multiple factors, however, it still shows considerable potential in terms of coverage with possibly feasible templates.